Research map / Vision-language and multimodal / Multimodal LLMs
CoLT: Teaching Multi-Modal Models to Think with Chain of Latent Thoughts
Lianyu Hu, shengqian qin, Zeqin Liao, Qing Guo, Liang Wan, Wei Feng, Yang Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace direct answering with three latent thought vectors | Qwen3-VL-8B · 8-benchmark average | accuracy: 69.5 -> 79.1 | Improved |
| replace textual CoT with latent thoughts | Qwen3-VL-8B · 8-benchmark average | accuracy: 75.7 -> 79.1 | Improved |
| replace SIM-CoT latent reasoning with CoLT | Qwen3-VL-8B · 8-benchmark average | accuracy: 74.0 -> 79.1 | Improved |
| add backward decoding loss to forward + internal losses (ablation) | CoLT · SeedBench, MMStar, AI2D, MMT | accuracy: 70.2 -> 74.8 | Improved |
| add forward decoding loss to backward + internal losses (ablation) | CoLT · SeedBench, MMStar, AI2D, MMT | accuracy: 71.3 -> 74.8 | Improved |
| add internal supervision loss to forward + backward losses (ablation) | CoLT · SeedBench, MMStar, AI2D, MMT | accuracy: 71.9 -> 74.8 | Improved |
| add forward decoding loss alone (ablation) | CoLT · SeedBench, MMStar, AI2D, MMT | accuracy: 63.6 -> 68.0 | Improved |
| increase latent steps K from 1 to 3 (ablation) | CoLT · SeedBench, MMStar, AI2D, MMT | accuracy: 68.9 -> 74.8 | Improved |
| increase latent steps K from 3 to 8 (ablation) | CoLT · SeedBench, MMStar, AI2D, MMT | accuracy: 74.8 -> 73.4 | Hurt |
| scale external decoder from Qwen3-0.6B to Qwen3-8B (ablation) | CoLT · SeedBench, MMStar, AI2D, MMT | accuracy: 74.8 -> 75.4 | No clear effect |
| train with K=3 and test at K=6 (ablation) | CoLT · SeedBench | accuracy: 77.5 -> 76.8 | No clear effect |
| replace text CoT with CoLT under 30% occlusion | Qwen3-VL-8B · MMStar occlusion noise | accuracy degradation: -13.8 -> -9.0 | Improved |
Closest papers by meaning
- SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.93)
- Multi-Modal Latent Space Learning for Chain-of-Thought Reasoning in Language Models (AAAI 2024 · similarity 0.92)
- ImgCoT: Compressing Long Chain of Thought into Compact Visual Tokens for Efficient Reasoning of Large Language Model (ICML 2026 · similarity 0.92)
- SIM-CoT: Supervised Implicit Chain-of-Thought (ICLR 2026 · similarity 0.92)
- Think in Latent, Explain in Language: Self-Explainable Latent Reasoning (ICML 2026 · similarity 0.91)
- Uni-CoT: Towards Unified Chain-of-Thought Reasoning Across Text and Vision (ICLR 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
